EP1469457A1 - Verfahren und System zur Vorverarbeitung von Sprachsignalen - Google Patents

Verfahren und System zur Vorverarbeitung von Sprachsignalen Download PDF

Info

Publication number
EP1469457A1
EP1469457A1 EP03007158A EP03007158A EP1469457A1 EP 1469457 A1 EP1469457 A1 EP 1469457A1 EP 03007158 A EP03007158 A EP 03007158A EP 03007158 A EP03007158 A EP 03007158A EP 1469457 A1 EP1469457 A1 EP 1469457A1
Authority
EP
European Patent Office
Prior art keywords
band
fbe
ssub
speech
full
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Withdrawn
Application number
EP03007158A
Other languages
English (en)
French (fr)
Inventor
Raquel Sony International Tato (Europe) GmbH
Thomas Sony International Kemp (Europe) GmbH
Antoni Sony International Abella (Europe) GmbH
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Sony Deutschland GmbH
Original Assignee
Sony International Europe GmbH
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Sony International Europe GmbH filed Critical Sony International Europe GmbH
Priority to EP03007158A priority Critical patent/EP1469457A1/de
Priority to JP2004078939A priority patent/JP2004341493A/ja
Priority to US10/809,162 priority patent/US7376559B2/en
Publication of EP1469457A1 publication Critical patent/EP1469457A1/de
Withdrawn legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS OR SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING; SPEECH OR AUDIO CODING OR DECODING
    • G10L15/00—Speech recognition
    • G10L15/02—Feature extraction for speech recognition; Selection of recognition unit
    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS OR SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING; SPEECH OR AUDIO CODING OR DECODING
    • G10L15/00—Speech recognition
    • G10L15/20—Speech recognition techniques specially adapted for robustness in adverse environments, e.g. in noise, of stress induced speech
    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS OR SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING; SPEECH OR AUDIO CODING OR DECODING
    • G10L21/00—Processing of the speech or voice signal to produce another audible or non-audible signal, e.g. visual or tactile, in order to modify its quality or its intelligibility
    • G10L21/02—Speech enhancement, e.g. noise reduction or echo cancellation

Definitions

  • the invention relates to a method for pre-processing speech, in particular to a method for recognizing speech.
  • the invention provides a method for pre-processing speech, in particular in a method for recognizing speech, according to claim 1.
  • the invention provides a speech pre-processing system, in particular integrated into a speech processing system, a computer program product, and a computer readable storage medium as defined in claims 14, 15, and 16, respectively. Further features and preferred embodiments are respectively defined in respective sub-claims and/or in the following description.
  • the method for pre-processing speech comprises the steps of receiving a speech signal, separating a spectrum of said speech signal into a given number of predetermined frequency sub-bands, analyzing said speech signal within each of said frequency sub-bands, thereby generating respective band-dependent acoustic feature data for each of said respective frequency sub-bands, which band-dependent acoustic feature data are at least in part representative for said speech signal with respect to a respective frequency sub-band, deriving band-dependent likelihoods for occurrences of speech elements or of sequences thereof within said speech signal based on said band-dependent acoustic feature data and/or a derivative thereof, analyzing said speech signal within said entire spectrum, thereby generating full-band acoustic feature data, which are at least in part representative for said speech signal with respect to said entire spectrum, deriving a full-band likelihood for occurrences of speech elements or of sequences thereof within said speech signal based on said full-band acous
  • a spectrum of a speech signal is divided into a given number of predetermined frequency sub-bands and acoustic feature data are extracted in each frequency sub-bands, which are then used to determine band-dependent likelihoods in each frequency band.
  • the spectrum of said speech signal is analyzed in the entire frequency band, thereby generating additional acoustic feature data, which are then used to derive a likelihood term, which depends on the entire spectrum.
  • the band-dependent likelihoods and the likelihood term are then combined to yield an overall likelihood by adding the different likelihood contributions of the band-dependent likelihoods and the likelihood term.
  • the overall likelihood will be dominated by either the band-dependent likelihoods or a combination thereof or by the likelihood term depending on the type of noise in the speech signal, i.e. the model will adopt automatically to the type of noise.
  • a broadband noise is resident in said speech signal, then the likelihood term will dominate the overall likelihood, because a broadband noise robust front-end may be used.
  • any front-end may be used, which is robust against broadband noise, e.g. a frequency filtering front end may be used for feature extraction and thus the broadband noise can effectively be dealt with.
  • narrow band noise resides in the speech signal the likelihood contribution to the overall likelihood of the band-dependent likelihoods will dominate. This is because, e.g. only one of said frequency sub-bands may be distorted, and the band-dependent likelihoods from undistorted frequency sub-bands are dominating.
  • said band-dependent likelihoods are combined to a union model likelihood by determining the number of uncorrupted frequency sub-bands of said frequency sub-bands, and adding all possible combinations of products of different band-dependent likelihoods corresponding to respective frequency sub-bands.
  • a certain number of corrupted frequency sub-bands is assumed, and the products depend on this number of assumed corrupted frequency sub-bands.
  • the number of factors in each addend is equal to said given number of predetermined frequency sub-bands minus the number of frequency sub-bands assumed to be corrupted, i.e. distorted.
  • the number of addends is equal to the number of frequency sub-bands, in the following equation denoted by N, over the number of frequency sub-bands minus the number of frequency sub-bands assumed to be distorted, in the following equation denoted by M, i.e. the number of addends is given by the binominal equation
  • the different products are then preferably added to yield said union model likelihood.
  • the union model likelihood will be dominated by the product containing band-dependent likelihoods, which are not or only a little distorted.
  • said band-dependent acoustic feature data comprise respective band-dependent mel-frequency cepstral coefficient features, which are based on mel-frequency cepstral coefficients and/or a derivative thereof derived from respective frequency sub-bands (F 1 , ..., F N ). This means, mel-frequency cepstral coefficient feature extraction is performed in each of said frequency sub-bands.
  • a predetermined broadband noise robustness technique is applied prior to deriving said likelihood term.
  • said full-band acoustic feature data may also comprise any other broadband noise robustness technique.
  • Said broadband noise robustness technique may be based on a frequency-filtering technique.
  • said broadband noise robustness technique may be based on a method of spectral-subtraction.
  • Said full-band acoustic feature data may comprise filter bank energy features, which are based on filter bank energies derived from said entire spectrum.
  • spectrum refers to a power density spectrum as typically used in speech recognition systems prior to deriving filter bank energies, i.e. using a mel-scaled filter bank. Also, the logarithm of the filter bank energies is taken as typical within a method for speech recognition.
  • said full-band acoustic feature data comprise filtered filter bank energy features, which are based on filtered filter bank energies derived from said entire spectrum. This way, the influence of broadband noise can be effectively reduced.
  • a frequency-filtering front end is applied to the full-band logarithmic filter bank energies. It should be noted, that all filter bank energies are logarithmic filter bank energies, hence, in the following the fact that the logarithm is taken may not be mentioned every time explicitly.
  • said full-band acoustic feature data comprise full-band mel frequency cepstral coefficient features, which are based on mel frequency cepstral coefficients and/or a derivative thereof derived from said entire spectrum. These features also contain information about the whole spectrum and can therefore advantageously be used, if no noise resides in said speech signal.
  • Said full-band acoustic feature data and/or said band-dependent acoustic feature data may also comprise PLP-linear prediction filter features, which are based on PLP-linear prediction filter coefficients.
  • PLP-linear prediction filter features which are based on PLP-linear prediction filter coefficients.
  • any other types of features can be applied, e.g. also logarithmic filter bank energy features directly or a derivative thereof, i.e. without filtering.
  • Typical derivatives may include cepstral features.
  • Said full-band acoustic feature data may also comprise spectrally changed full-band mel-frequency cepstral coefficient features, which are generated by applying a method of spectral subtraction to said full-band mel-frequency cepstral coefficient features. Therefore, a method of noise estimation is performed detecting the noise resident in said speech signal. This technique also effectively deals with broadband noise.
  • said band-dependent likelihoods and said likelihood term may be determined using a probability estimator. It is possible, that the probability estimator is trained with data containing speech corrupted by various types of noise.
  • said filtered filter bank energies are derived from said filter bank energies by subtracting a first filter bank energy from a second filter bank energy, wherein said first filter bank energy corresponds to a first discrete frequency and said second filter bank energy corresponds to a second discrete frequency, lying two discrete frequency steps after said first filter bank energy.
  • a speech pre-processing system which is in particular integrated into a speech processing system, is capable of performing or realizing a method for pre-processing speech as defined above and/or the steps thereof.
  • a computer program product comprises a computer program means adapted to perform and/or to realize the method of pre-processing speech and/or the steps thereof, when it is executed on a computer, a digital signal processing means, and/or the like.
  • a computer readable storage medium comprises a computer program product as defined above.
  • filter bank energies log FBE are derived from the entire band, i.e. for the entire spectrum. Then, the filter bank energies log FBE from the entire band are separated into a predetermined number of frequency sub-bands, here, a first frequency sub-band F 1 , a second frequency sub-band F 2 , and a third frequency sub-band F 3 . It should be noted that these are logarithmic filter bank energies log FBE as commonly used.
  • the results are first sub-band filter bank energies FBE-F 1 , second sub-band filter bank energies FBE-F 2 , and third sub-band filter bank energies FBE-F 3 .
  • filter bank energies mel-frequency cepstral coefficients are derived, i.e. a MFCC feature extraction is performed for each of the frequency sub-bands, i.e. the first sub-band filter bank energies FBE-F 1 , the second sub-band filter bank energies FBE-F 2 , and the third sub-band filter bank energies FBE-F 3 .
  • first sub-band acoustic feature data O 1 are derived from said first sub-band filter bank energies FBE-F 1
  • second sub-band acoustic feature data O 2 are derived from said second sub-band filter bank energies FBE-F 2
  • third sub-band acoustic feature data O 3 are derived from said third sub-band filter bank energies FBE-F 3 .
  • the derived sub-band acoustic feature data are then used as input for a probability estimator PE, which estimates band-dependent likelihoods within each sub-band acoustic feature data for speech elements corresponding to the speech signal.
  • Speech elements may e.g. be words, phones, or sub-phonetic units.
  • the probability estimator PE estimates a first band-dependent likelihood b 1 from the first sub-band acoustic feature data O 1 , further a second band-dependent likelihood b 2 from the second sub-band acoustic feature data O 2 , and a third band-dependent likelihood b 3 from the third sub-band acoustic feature data O 3 .
  • these band-dependent likelihoods are combined to a union model likelihood B U.MFCC .
  • the formula for calculating the union model likelihood B U.MFCC differs depending on the number M of frequency sub-bands assumed to be distorted:
  • the union model likelihood B U.MFCC can be written as In the example of Fig. 1 the union model likelihood B U.MFCC is given by equation (2) as explained above, i.e. equation (3) simplifies to equation (2).
  • part A which is denoted as such in Fig. 1.
  • a first embodiment to determine a frequency filter likelihood term B FF is depicted.
  • a logarithmic filter bank energy feature extraction log FBE is performed with respect to the entire spectrum F of the speech signal S.
  • the result are full-band filter bank energy features FBE-F of the entire spectrum F.
  • the full-band filter bank energy features FBE-F are subjected to a broadband noise robust front-end.
  • any broadband noise robust front-end i.e. any front-end which is robust against broadband noise, may be used. It may also be possible that no broadband noise robust front-end is used.
  • a frequency-filtering front-end is used, which is described in detail in prior art document "Time and frequency filtering of filter-bank energies for robust HMM speech recognition" by C. Nadeu, D. Macho, and J. Hernando, Speech Communication, Vol. 34, Issue 1-2, April 2001, pp 93-114. As far as frequency filtering is concerned, please see this prior art document, the content of which is included herein by reference.
  • Eq. (4) means, that in order to obtain a filtered filter bank energy value of said filtered filter bank energy features FFBE at a certain discrete frequency i, the value of the previous filter bank energy FBE i-1 corresponding to f(i-1) in eq. (4) is subtracted from the following filter bank energy FBE i+1 corresponding to f(i+1) in eq. (4).
  • the values of variable i in eq. (4) correspond to discrete frequencies as indicated in Fig. 3.
  • the values of f(i+1) and f(i-1) correspond to coefficients of the corresponding Fourier transformation. This means, frequency filtering is done for each feature vector, independently of surrounding feature vectors. Feature vectors are thereby extracted every 10ms from the speech signal S.
  • a filtered filter bank energy value FFBE i at a certain frequency i may be calculated by subtracting a filter bank energy value FBE i-1 corresponding to a previous frequency i-1 from a filter bank energy value FBE i+1 corresponding to a following frequency i+1.
  • the filtered filter bank energy features FFBE are then used as an input to a probability estimator PE, which estimates the frequency filtered likelihood term B FF .
  • a second embodiment B 2 to determine the overall likelihood is depicted.
  • a spectral subtraction likelihood term B SSUB is combined with the union model likelihood B U.MFCC.
  • this spectral subtraction likelihood term B SSUB For determining this spectral subtraction likelihood term B SSUB , first, the noise within the power density spectrum PDS, is determined.
  • the power density spectrum PDS is derived from the speech signal S.
  • the result of the noise estimation is estimated noise EN.
  • the method of spectral subtraction SSUB is applied, which uses as input the estimated noise EN and the power density spectrum PDS.
  • the output of the method of spectral subtraction SSUB are power density spectrum spectral subtraction features PDS-SSUB.
  • These power density spectrum spectral subtraction features PDS-SSUB are subjected to a logarithmic filter bank energy feature extraction log FBE. Thereby, full-band spectrally subtracted filter bank energies FBE-F-SSUB are derived. These full-band spectrally subtracted filter bank energies FBE-F-SSUB are subjected to a mel-frequency cepstral coefficient MFCC feature extraction, wherein spectrally-changed full-band mel-frequency cepstral coefficient features O F.SSUB are generated. These spectrally-changed full-band mel-frequency cepstral coefficient features O F.SSUB , are then used by the probability estimator PE to estimate the spectral subtraction likelihood term B SSUB .
  • Fig. 3 depicts the filter bank energy feature extraction log FBE.
  • the basis forms the power density spectrum PDS.
  • the frequency range of the power density spectrum PDS is separated into overlapping intervals, wherein a weighting function is assigned to each interval.
  • the intervals are chosen according to the resolution of the human ear, i.e. a high density of intervals occurs within sensitive regions of the human ear, wherein a low density of intervals occurs within less sensitive regions of the human ear.
  • the invention performs an integration of broadband noise cancellation techniques into the Union Model approach.
  • the Union Model approach is a powerful technique for dealing with narrow band noise, including non-stationary noises. However it is not very well suited for stationary broadband noises, where traditional techniques seem to perform better. Integration of such techniques into the Union Model concept will allow dealing with any kind of noise in any circumstances.
  • the Frequency Filtering front-end applied to full-band based speech recognition achieve better results than the cepstral coefficients for speech corrupted by various types of wide-band real-world noises.
  • the Frequency Filtering front-end is integrated as an additional factor in the formula of the output probabilities calculation from the Union Model.
  • the main purpose of this approach is to have a model that will use MFCC in case of narrow band noise, and Frequency Filtering for broadband noise, selecting automatically the most appropriate from the output probability calculation.
  • N for MFCC front-end (to calculate feature vectors for each subband) and 1 for Frequency Filtering approach, but applied to full-band (to calculate one single feature vector for the whole band): B(O 1 ) ⁇ b MFCC 1 b MFCC 2 + b MFCC 1 b MFCC 3 + b MFCC 2 b MFCC 3 + B FF
  • N for MFCC front-end (to calculate feature vectors for each sub-band) and 1 for MFCC+Spectral Substraction applied to full-band (to calculate one single feature vector for the whole band):
  • B ( O t ) ⁇ b MFCC 1 b MFCC 2 + b MFCC 1 b MFCC 3 + b MFCC 2 b MFCC 3 + B MFCC+SS
  • the invention introduces an extension of the union model approach that allows robustness against broadband noise.
  • a number of recent studies reveal that union model approach with Mel-Frequency Cepstral Coefficients (MFCCs) as front-end, offers robustness to band limited corruption, without requiring information about the noise.
  • MFCCs Mel-Frequency Cepstral Coefficients
  • frequency filtering front-end applied to full-band based speech recognition is proved to achieve better results than cepstral coefficients for speech corrupted by various types of broadband real-world noises.
  • frequency filtering front-end on full-band is integrated as an additional stream in the union model.
  • the double extraction of information i.e. different sets of features from the same frequency band, can be seen as a diversity technique, which results in a more robust system.
  • frequency filtering front end for the full-band approach has proved to be a clear alternative to the cepstral coefficients for speech recognition in presence of unknown broadband noise. From a robust speech recognition point of view, it will be desirable to have a system, which is able to deal with as many types of noise as possible.
  • a combination of the probabilistic union model and the frequency filtering technique is known, which has the advantage of dealing with both frequency localized noise and wide band noise, but only under very specific circumstances as we will see later. At this point, it seems that the problem to find a technique that can deal with broadband and band limited noise is still to be solved.
  • the invention consist in integrating frequency filtering front-end applied to full-band, as an additional stream in the union model, being represented in the sum of the output probability calculation as an independent term.
  • the goal is to have a model that will use cepstral coefficients combined with subband approach in case of narrow-band noise, and frequency filtering with full-band approach for broadband noise, selecting automatically the most appropriate from the output probability calculation.
  • cepstral coefficients combined with subband approach in case of narrow-band noise, and frequency filtering with full-band approach for broadband noise, selecting automatically the most appropriate from the output probability calculation.
  • the union model is known for likelihood combination in the presence of band limited additive noise. Essentially, the signal is split up in N frequency bands. Under the assumption that M (M ⁇ N) bands are distorted, the likelihood, i.e. total output probability, can be computed as the sum of the likelihood contributions of all N-M bands combinations. The principal idea is that if a combination includes the corrupted band, then its likelihood is very low, and therefore the sum of the individual likelihood contributions is dominated by the one combination of bands where the noisy band is excluded. The interesting property of the union model is now that it is not necessary to know which of the bands is corrupted.
  • the frequency filtering front-end has been successfully used as an alternative to the cepstral coefficients for noisy speech recognition.
  • the idea is to generate a set of feature parameters by filtering the logarithm filter bank energies (FBE), with an effect of decorrelation and the additional advantage of lying in the frequency domain.
  • the FIR filter with transfer function z-z -1 is applied to 15 logarithmic FBE, and delta parameters are calculated, resulting in a 45 feature vector.
  • Table 1 presents the experiments carried out for the noise conditions mentioned earlier, including clean speech. In clean conditions, the results are comparable to the full-band approach with MFCC.
  • Frequency filtering success is based, to some extent, on its ability for noise cancellation between the filter banks, based on the assumption that the noise is stationary in the frequency domain, which is not the case for narrow-band noises.
  • the frequency filtering is used to produce a 9 feature vector for each of the 5 subband in the probabilistic union model, instead of MFCC.
  • This aims to benefit from the good noise localization capability that offers this technique, which should isolate the noise in the corresponding subbands, leaving the others unaffected, in case of narrow-band noise.
  • it should provide robustness against broadband noise.
  • the invention uses an additional stream for Frequency Filtering full band front end, which will be explained more detailed in the following.
  • the union model approach with MFCC as front end overcomes the signal quality deterioration by the assumption of band limited aditive noise, and by effectively ignoring the contribution of the distorted signal band in the likelihood computation.
  • the frequency filtering front end applied to full-band based speech recognition achieves better results than the cepstral coefficients for speech corrupted by various types of wide-band real-world noises.
  • frequency filtering front-end is integrated as an additional factor in the formula of the output probabilities calculation from the union model, i.e. integrated as an independent stream in the union model.
  • the new output probabilty will look like: B ( o t ) ⁇ b 1 b 2 b 3 b MFCC 4 + ... + b 2 b 3 b 4 b 5 MFCC + B Freq.Filt.
  • each observation vector at time t can be split into 6 independent data streams, and the formula for computing the output distributions in the Baum-Welch Re-estimation algorithm for the problem of parameter estimation can be written as, where there are 2 mixtures components in each stream, c sm is the weight of the m'th component, and N(.; ⁇ , ⁇ ) is a multivariate Gaussian with mean vector ⁇ and covariance matrix ⁇ .
  • the weight of each of the 6 streams is set to 1.
  • the feature vector for each of the 5 subband streams is composed of 4 MFCC and 4 ⁇ MFCC, i.e. 8 features each.
  • the fullband stream contains a 45 feature vector, resulting from the application of FIR filter with transfer function z-z -1 to 15 logarithmic FBE, and the estimation of delta parameters.
  • Results from the fifth column in Table 1 reproduce the evaluation of this method for all the noises that are object of an investigation.
  • the recognition rates are even better that any of the individual systems.
  • the band limited noise e.g. clock
  • the accuracy with frequency filtering front-end and full-band is very low, 20,7%, while with the union model and MFCC front end it is quite high, 92,2%.
  • the new model outperforms the best result with a recognition rate of 94,9%. The same behavior holds also for music and broadband noise.
  • the order of the union model i.e. the number of assumed noisy bands, M, that gives better performance, increases when such model is integrated with a frequency filtering full-band stream.
  • M the number of assumed noisy bands
  • the system already has some broadband information in the sixth stream, especially information related to the joint probability distribution, and therefore it is preferable to avoid as many noisy subbands as possible.

Landscapes

  • Engineering & Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Computational Linguistics (AREA)
  • Health & Medical Sciences (AREA)
  • Audiology, Speech & Language Pathology (AREA)
  • Human Computer Interaction (AREA)
  • Physics & Mathematics (AREA)
  • Acoustics & Sound (AREA)
  • Multimedia (AREA)
  • Circuit For Audible Band Transducer (AREA)
  • Compression, Expansion, Code Conversion, And Decoders (AREA)
EP03007158A 2003-03-28 2003-03-28 Verfahren und System zur Vorverarbeitung von Sprachsignalen Withdrawn EP1469457A1 (de)

Priority Applications (3)

Application Number Priority Date Filing Date Title
EP03007158A EP1469457A1 (de) 2003-03-28 2003-03-28 Verfahren und System zur Vorverarbeitung von Sprachsignalen
JP2004078939A JP2004341493A (ja) 2003-03-28 2004-03-18 音声前処理方法
US10/809,162 US7376559B2 (en) 2003-03-28 2004-03-25 Pre-processing speech for speech recognition

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
EP03007158A EP1469457A1 (de) 2003-03-28 2003-03-28 Verfahren und System zur Vorverarbeitung von Sprachsignalen

Publications (1)

Publication Number Publication Date
EP1469457A1 true EP1469457A1 (de) 2004-10-20

Family

ID=32892862

Family Applications (1)

Application Number Title Priority Date Filing Date
EP03007158A Withdrawn EP1469457A1 (de) 2003-03-28 2003-03-28 Verfahren und System zur Vorverarbeitung von Sprachsignalen

Country Status (3)

Country Link
US (1) US7376559B2 (de)
EP (1) EP1469457A1 (de)
JP (1) JP2004341493A (de)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112750454A (zh) * 2020-07-16 2021-05-04 鸣飞伟业技术有限公司 一种基于应急通信后端箱的应用系统

Families Citing this family (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US8175877B2 (en) * 2005-02-02 2012-05-08 At&T Intellectual Property Ii, L.P. Method and apparatus for predicting word accuracy in automatic speech recognition systems
US20070250311A1 (en) * 2006-04-25 2007-10-25 Glen Shires Method and apparatus for automatic adjustment of play speed of audio data
DE102009035524B4 (de) * 2009-07-31 2012-11-29 Fraunhofer-Gesellschaft zur Förderung der angewandten Forschung e.V. Verfahren zur Erkennung eines oder mehrerer Nutzsignale innerhalb eines Quellsignals
CN101996628A (zh) * 2009-08-21 2011-03-30 索尼株式会社 提取语音信号的韵律特征的方法和装置
AU2015271665B2 (en) 2014-06-05 2020-09-10 Interdev Technologies Systems and methods of interpreting speech data
US9812128B2 (en) * 2014-10-09 2017-11-07 Google Inc. Device leadership negotiation among voice interface devices
US10026395B1 (en) * 2017-01-06 2018-07-17 Applied Brain Research Inc. Methods and systems for extracting auditory features with neural networks
EP3564949A1 (de) * 2018-04-23 2019-11-06 Spotify AB Aktivierungsauslöserverarbeitung
KR102241364B1 (ko) * 2018-05-29 2021-04-16 연세대학교 산학협력단 음성 신호를 이용한 사용자 스트레스 판별 장치 및 방법
US11947593B2 (en) 2018-09-28 2024-04-02 Sony Interactive Entertainment Inc. Sound categorization system
CN110197670B (zh) * 2019-06-04 2022-06-07 大众问问(北京)信息科技有限公司 音频降噪方法、装置及电子设备

Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2002095730A1 (en) * 2001-05-21 2002-11-28 Queen's University Of Belfast Interpretation of features for signal processing and pattern recognition

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
DE4316297C1 (de) * 1993-05-14 1994-04-07 Fraunhofer Ges Forschung Frequenzanalyseverfahren
JP2655046B2 (ja) * 1993-09-13 1997-09-17 日本電気株式会社 ベクトル量子化装置
US5625743A (en) * 1994-10-07 1997-04-29 Motorola, Inc. Determining a masking level for a subband in a subband audio encoder

Patent Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2002095730A1 (en) * 2001-05-21 2002-11-28 Queen's University Of Belfast Interpretation of features for signal processing and pattern recognition

Non-Patent Citations (5)

* Cited by examiner, † Cited by third party
Title
DATABASE INSPEC [online] THE INSTITUTION OF ELECTRICAL ENGINEERS, STEVENAGE, GB; JIANG WEN-JIAN ET AL: "A differential hybrid subband method for noisy speech recognition", XP002251485, Database accession no. 7406613 *
JI MING ET AL: "Union: a new approach for combining sub-band observations for noisy speech recognition", SPEECH COMMUNICATION, AMSTERDAM, NL, vol. 34, no. 1-2, 2001, pages 41 - 55, XP002209287, ISSN: 0167-6393 *
JOURNAL OF CHINA INSTITUTE OF COMMUNICATIONS, JAN. 2002, PEOPLE'S POSTS & TELECOMMUN. PUBLISHING HOUSE, CHINA, vol. 23, no. 1, pages 18 - 24 *
NADEU C ET AL: "Time and frequency filtering of filter-bank energies for robust HMM speech recognition", SPEECH COMMUNICATION, APRIL 2001, ELSEVIER, NETHERLANDS, vol. 34, no. 1-2, pages 93 - 114, XP002251486, ISSN: 0167-6393 *
TIBREWALA S ET AL: "Sub-band based recognition of noisy speech", 1997 IEEE INTERNATIONAL CONFERENCE ON ACOUSTICS, SPEECH, AND SIGNAL PROCESSING (CAT. NO.97CB36052), 1997 IEEE INTERNATIONAL CONFERENCE ON ACOUSTICS, SPEECH, AND SIGNAL PROCESSING, MUNICH, GERMANY, 21-24 APRIL 1997, 1997, Los Alamitos, CA, USA, IEEE Comput. Soc. Press, USA, pages 1255 - 1258 vol.2, XP010226029, ISBN: 0-8186-7919-0 *

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112750454A (zh) * 2020-07-16 2021-05-04 鸣飞伟业技术有限公司 一种基于应急通信后端箱的应用系统
CN112750454B (zh) * 2020-07-16 2022-08-09 鸣飞伟业技术有限公司 一种基于应急通信后端箱的应用系统

Also Published As

Publication number Publication date
JP2004341493A (ja) 2004-12-02
US7376559B2 (en) 2008-05-20
US20040236570A1 (en) 2004-11-25

Similar Documents

Publication Publication Date Title
Burshtein et al. Speech enhancement using a mixture-maximum model
Xiao et al. Normalization of the speech modulation spectra for robust speech recognition
Bahoura et al. Wavelet speech enhancement based on time–scale adaptation
US20090163168A1 (en) Efficient initialization of iterative parameter estimation
Visser et al. A spatio-temporal speech enhancement scheme for robust speech recognition in noisy environments
JPH07271394A (ja) 確実な電話音声認識のための信号バイアスの除去
US7376559B2 (en) Pre-processing speech for speech recognition
Hung et al. On the use of weighted filter bank analysis for the derivation of robust MFCCs
US6377918B1 (en) Speech analysis using multiple noise compensation
US20030187637A1 (en) Automatic feature compensation based on decomposition of speech and noise
Gupta et al. Speech enhancement using MMSE estimation and spectral subtraction methods
Pattanayak et al. Pitch-robust acoustic feature using single frequency filtering for children’s KWS
de Veth et al. Acoustic backing-off as an implementation of missing feature theory
Joshi et al. Sub-band based histogram equalization in cepstral domain for speech recognition
You et al. Kalman filtering speech enhancement incorporating masking properties for mobile communication in a car environment
Hernando Pericás et al. CDHMM speaker recognition by means of frequency filtering of filter-bank energies
Korba et al. Robust speech recognition using perceptual wavelet denoising and mel-frequency product spectrum cepstral coefficient features
Zhu et al. Maximum likelihood sub-band adaptation for robust speech recognition
Vanajakshi et al. Modified Wiener filter based speech restoration with autocorrelation of isolated Kannada word speech recognition
Das et al. Vector taylor series expansion with auditory masking for noise robust speech recognition
Acero et al. Towards environment-independent spoken language systems
Seltzer et al. Training wideband acoustic models using mixed-bandwidth training data via feature bandwidth extension
Ming et al. A comparative study of methods for handheld speaker verification in realistic noisy conditions
Mwema et al. A spectral subtraction method for noise reduction in speech signals
Wolfel A joint particle filter and multi-step linear prediction framework to provide enhanced speech features prior to automatic recognition

Legal Events

Date Code Title Description
PUAI Public reference made under article 153(3) epc to a published international application that has entered the european phase

Free format text: ORIGINAL CODE: 0009012

AK Designated contracting states

Kind code of ref document: A1

Designated state(s): AT BE BG CH CY CZ DE DK EE ES FI FR GB GR HU IE IT LI LU MC NL PT SE SI SK TR

AX Request for extension of the european patent

Extension state: AL LT LV MK RO

17P Request for examination filed

Effective date: 20041124

17Q First examination report despatched

Effective date: 20050428

AKX Designation fees paid

Designated state(s): DE FR GB

RAP1 Party data changed (applicant data changed or rights of an application transferred)

Owner name: SONY DEUTSCHLAND GMBH

17Q First examination report despatched

Effective date: 20050428

RAP1 Party data changed (applicant data changed or rights of an application transferred)

Owner name: SONY DEUTSCHLAND GMBH

STAA Information on the status of an ep patent application or granted ep patent

Free format text: STATUS: THE APPLICATION IS DEEMED TO BE WITHDRAWN

18D Application deemed to be withdrawn

Effective date: 20070724